happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

metal : adjust extra size for FA buffer to avoid reallocations (#18545)

This commit is contained in:
Georgi Gerganov 2026-01-02 19:02:18 +02:00 committed by GitHub
parent af1e8e1a6c
commit f38de16341
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-metal/ggml-metal-ops.cpp
View File

@ -2181,7 +2181,11 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
const bool has_mask = op->src[3] != nullptr;
if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
// note: the non-vec kernel requires more extra memory, so always reserve for it
GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG);
//if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
if (false) {
// note: always reserve the padding space to avoid graph reallocations
//const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
const bool has_kvpad = true;